task-digestion

Provide the ticket key or URL and the Skill fetches the mission, decomposes it into atomic requirements, groups them into tasks, and identifies real dependencies. The result is a task-plan.json DAG with levels showing which tasks can run in parallel.

A dependency exists only when one task genuinely cannot start or be verified until another finishes, such as a shared interface, schema, or sequential input-output contract. Topical proximity, convenience ordering, or being in the same milestone do not count.

The plan is emitted as task-plan.json, a DAG with stable task ids, priority fields, depends_on edges, and derived levels, plus a TASK_PLAN.md markdown companion rendering the same data in tables grouped by level.

Yes. Tasks sharing a level have no dependency relationship in either direction, so they can be assigned to separate agents or engineers concurrently. A task only starts once every id in its depends_on list is done.

A cycle signals a design problem in the mission itself, so the Skill stops rather than forcing an arbitrary order. It names the exact task ids involved and puts the cycle in open questions for a human to resolve, usually by merging tasks or picking an interface owner.

Skip it for trivial missions with a single task and no dependency question; the Skill will say so and emit a one-node plan. It also does not cover implementation, code review, or testing, which belong to other skills.